The Trillion-Dollar Circle

In one minute:

Anthropic's private valuation rose from $380bn to $965bn in 105 days, a factor of 2.54, and OpenAI's reported talks put it at about 1.64 times its March mark.

The same few firms now sell to, guarantee and lease from one another, so a valuation depends partly on a loop of their own orders.

Forecasters should price concentration, not just growth. Our registered forecasts F10 (35%) and F11 (85%) will be scored.

1 August 2026, Nairobi

On the first morning of August, a kiosk owner in Nairobi could take Pochi la Biashara payments of up to KSh 200 without paying a fee. The change was announced by Safaricom. It was small, and it lived inside a phone. It was also a reminder that a great deal of the city's money runs over a few rails owned by a few firms.

That same month, the largest money in the technology world was moving in a loop.

What was visible

In February, the signal was already on the record. Private AI valuations were climbing in large steps, with each round priced well above the last. A forecaster reading the rounds in order would have seen the marks rising quickly, and would have asked a sharper question than "how high can they go?": who is paying whom, and how much of that money comes back to the seller?

In hindsight, the arithmetic was visible well before August. We are not claiming anyone predicted the August details. We are showing what a careful reader could have counted.

What happened

Table sources: $380bn, $852bn, $965bn, $105bn, about 2 million chips, at about $1.4tn.

The Ohio guarantee is the thread to follow. Nvidia's guarantee sits on the PORTS campus in Pike County, where OpenAI holds 20-year leases for about 4.25 GW of IT load across nine data centres. The chipmaker is standing behind the buyer of its own product. In the same weeks, Nvidia reported second-quarter fiscal 2027 revenue of $96.2bn, with data-centre revenue of $89.0bn, up 117% year on year. Its supply and manufacturing commitments rose to $279bn, from $119bn the quarter before.

Outside the loop, the money was thinning. Global venture funding in August fell 25% from July to about $42bn, reported from Crunchbase data.

Not every mark was a closed deal. The coding-agent company Cognition was in early talks at a valuation of at least $40bn, up more than 50% from $26bn earlier in 2026. Talks are not rounds, and we count them separately.

The math

The concept is a growth factor and a clock. Divide the later mark by the earlier one, then count the days between them.

growth factor = later valuation ÷ earlier valuation days = calendar days between the two marks

Step 1, Anthropic. From 12 February to 28 May is 16 days left in February, plus 31 in March, 30 in April and 28 in May, which makes 105 days.

965 ÷ 380 = 2.54 in 105 days

Step 2, OpenAI. Measured from its 31 March mark to the reported talks at about $1.4tn, the factor is:

1,400 ÷ 852 = 1.64

The OpenAI figure is talks, not a closed round. We use it because it is the most recent public mark, and we flag its status.

Step 3, what the number cannot tell you. A growth factor measures speed, not soundness. A valuation can rise fast because revenue is rising, or because the buyers and sellers are the same firms. The arithmetic here cannot separate those two. Part of the problem is that the sources mix different kinds of commitment. A round that is still in talks, a guarantee that is capped, and a chip order for 2027 and 2028 are not the same kind of money, and a reader should not add them up as if they were. Keep them apart, and the loop is easier to see. What it can show is that the loop widened while the marks climbed. The market noticed. The S&P 500 fell in five of six sessions after its high, and the Nasdaq Composite fell 1.88% on 14 October.

The Pawanax lens

Concentration becomes practical when it reaches a clinic's bill. Anthropic made its Claude Sonnet 5 introductory price of $2 per million input tokens and $10 per million output tokens permanent, cancelling a planned increase. That was good news for any clinic that had budgeted for it. It was also a reminder that one company's decision changed the cost of a tool that many clinics now rely on. Google's launch-tier price for Gemini 4 was reported as temporary, which is the same lesson in a different form.

For a clinic in Nairobi, Kisumu or Kilifi, the practical rule is simple. Plan for a price that can change by one firm's decision, and keep a second route for the work that cannot stop. At Pawanax, we treat supplier concentration as a design constraint from the first line of code, not a footnote for the investor deck.

The next 90 days

We have registered two forecasts on this loop. Both are scored with the Brier score, the squared gap between a probability and the outcome, where the outcome is 1 for YES and 0 for NO. A forecaster who answers 50% on everything scores 0.25, so we claim skill only if we beat that clearly.

• F11: 85%. At least two of OpenAI, Anthropic and Google release a new flagship model generation before 31 January 2027. The release cadence is already fast: GPT-6 Astra arrived on 3 September, and Google's Gemini 4 was reported in broad release in mid-October. We score it on 31 January 2027.
• F10: 35%. The Nasdaq Composite closes at least 15% below its 2026 high on any day before 16 April 2027. We score it on closing prices, not on headlines. We think this is the riskiest gap on the board, because the loop depends on valuations that rates can reprice.

The loop is real, and so is the arithmetic. We will publish the scores whichever way they fall.

Part 8 of The Signal Ledger. Read the series: [The Signal Ledger](/blog/signal-ledger-2026-read-backwards).

Cover image: Racks Amravati Data Center.jpg — PiDatacenters · CC BY-SA 4.0. Photographed 17 August 2021 at Pi Data Centers, Amravati, India. Illustrative data center infrastructure, not a specific 2026 facility or event. Resized for web display.